CS231n_8_9

Lec8

Spatial Localization and Detection

Tasks

Classification, Classificatin+Localization, Object Detection, Instance Segmentation.

Localization, I:Image O:Box in the image(x,y,w,h) E:Intersection over Union.(IoU)

Claissification+Localization

Output: Single Label and Bounding Box

Idea one: Two task head
1.Train a CNN
2.Attach new fully-connected “regression head” to the network(FC)
2.1 Classification Head
2.2 Regression Head
3.Train the regression head only with SGD and L2 loss
4.At test time use both heads

Idea two: Sliding Window
Input: Bounding Box
Iteratively refine the BB into a optimal size and place.

Objects Detection

Output: all the exist labels and BBs

Problem: Need to test many positions and scales, use computationally demanding classifier

Solution: Only look at a tiny subset of possible positions

Region Proposal:Bottom-up segmentation.

RCNN: 1. Train a classification model on ImageNet

  1. Fine-tune model for detection(Throw finaly FC rathter than 20 Objects and one background, that is 4096*21 for the last layer.)
  2. Extract Features: Extract region proposals for all images, save the pool5 features to disk.
  3. Train one binary SVM per class to claissify region features.
  4. Bbox regression: Train a linear model to fine-grain the bbox

Fast-RCNN: Build an end-to-end framework, much faster than RCNN.

Summary

Lec9 Understanding and Visualizing

Visualizing the weights, t-SNE visualization

Deconv Approaches:
1.Feed image into net
2.pick a layer, set gradients of the score vector to [0 0 1 .. 0], then bp to image
3.Do a small “Image Update”
4.Forward the Image

  1. Go to step 2

Deconv

Learn to visualize the weights, also deconv to reconstruct an larger size output.

Deconv: reverse the convolution filter
DePool: record the position and set other be zero.
DeReLU: The same as the ReLU.

Neural Style

  1. extract content targets
  2. extract style targets
  3. Optimize over image

BackPropping is powerful

  • Understanding
  • Segmenting Objects in the Image
  • Inverting codes and introducing privacy concerns
  • Fun(NeuralStyle/DeepDream)
  • Confusion and chaos(Adversarial Examples)
坚持分享,支持原创